source: sasmodels/sasmodels/kernelcl.py @ 01c8d9e

core_shell_microgelsmagnetic_modelticket-1257-vesicle-productticket_1156ticket_1265_superballticket_822_more_unit_tests
Last change on this file since 01c8d9e was 01c8d9e, checked in by Suczewski <ges3@…>, 11 months ago

beta approximation, first pass

  • Property mode set to 100644
File size: 24.0 KB
Line 
1"""
2GPU driver for C kernels
3
4There should be a single GPU environment running on the system.  This
5environment is constructed on the first call to :func:`env`, and the
6same environment is returned on each call.
7
8After retrieving the environment, the next step is to create the kernel.
9This is done with a call to :meth:`GpuEnvironment.make_kernel`, which
10returns the type of data used by the kernel.
11
12Next a :class:`GpuData` object should be created with the correct kind
13of data.  This data object can be used by multiple kernels, for example,
14if the target model is a weighted sum of multiple kernels.  The data
15should include any extra evaluation points required to compute the proper
16data smearing.  This need not match the square grid for 2D data if there
17is an index saying which q points are active.
18
19Together the GpuData, the program, and a device form a :class:`GpuKernel`.
20This kernel is used during fitting, receiving new sets of parameters and
21evaluating them.  The output value is stored in an output buffer on the
22devices, where it can be combined with other structure factors and form
23factors and have instrumental resolution effects applied.
24
25In order to use OpenCL for your models, you will need OpenCL drivers for
26your machine.  These should be available from your graphics card vendor.
27Intel provides OpenCL drivers for CPUs as well as their integrated HD
28graphics chipsets.  AMD also provides drivers for Intel CPUs, but as of
29this writing the performance is lacking compared to the Intel drivers.
30NVidia combines drivers for CUDA and OpenCL in one package.  The result
31is a bit messy if you have multiple drivers installed.  You can see which
32drivers are available by starting python and running:
33
34    import pyopencl as cl
35    cl.create_some_context(interactive=True)
36
37Once you have done that, it will show the available drivers which you
38can select.  It will then tell you that you can use these drivers
39automatically by setting the SAS_OPENCL environment variable, which is
40PYOPENCL_CTX equivalent but not conflicting with other pyopnecl programs.
41
42Some graphics cards have multiple devices on the same card.  You cannot
43yet use both of them concurrently to evaluate models, but you can run
44the program twice using a different device for each session.
45
46OpenCL kernels are compiled when needed by the device driver.  Some
47drivers produce compiler output even when there is no error.  You
48can see the output by setting PYOPENCL_COMPILER_OUTPUT=1.  It should be
49harmless, albeit annoying.
50"""
51from __future__ import print_function
52
53import os
54import warnings
55import logging
56import time
57
58import numpy as np  # type: ignore
59
60
61# Attempt to setup opencl. This may fail if the opencl package is not
62# installed or if it is installed but there are no devices available.
63try:
64    import pyopencl as cl  # type: ignore
65    from pyopencl import mem_flags as mf
66    from pyopencl.characterize import get_fast_inaccurate_build_options
67    # Ask OpenCL for the default context so that we know that one exists
68    cl.create_some_context(interactive=False)
69    HAVE_OPENCL = True
70    OPENCL_ERROR = ""
71except Exception as exc:
72    HAVE_OPENCL = False
73    OPENCL_ERROR = str(exc)
74
75from . import generate
76from .kernel import KernelModel, Kernel
77
78# pylint: disable=unused-import
79try:
80    from typing import Tuple, Callable, Any
81    from .modelinfo import ModelInfo
82    from .details import CallDetails
83except ImportError:
84    pass
85# pylint: enable=unused-import
86
87# CRUFT: pyopencl < 2017.1  (as of June 2016 needs quotes around include path)
88def quote_path(v):
89    """
90    Quote the path if it is not already quoted.
91
92    If v starts with '-', then assume that it is a -I option or similar
93    and do not quote it.  This is fragile:  -Ipath with space needs to
94    be quoted.
95    """
96    return '"'+v+'"' if v and ' ' in v and not v[0] in "\"'-" else v
97
98def fix_pyopencl_include():
99    """
100    Monkey patch pyopencl to allow spaces in include file path.
101    """
102    import pyopencl as cl
103    if hasattr(cl, '_DEFAULT_INCLUDE_OPTIONS'):
104        cl._DEFAULT_INCLUDE_OPTIONS = [quote_path(v) for v in cl._DEFAULT_INCLUDE_OPTIONS]
105
106if HAVE_OPENCL:
107    fix_pyopencl_include()
108
109# The max loops number is limited by the amount of local memory available
110# on the device.  You don't want to make this value too big because it will
111# waste resources, nor too small because it may interfere with users trying
112# to do their polydispersity calculations.  A value of 1024 should be much
113# larger than necessary given that cost grows as npts^k where k is the number
114# of polydisperse parameters.
115MAX_LOOPS = 2048
116
117
118# Pragmas for enable OpenCL features.  Be sure to protect them so that they
119# still compile even if OpenCL is not present.
120_F16_PRAGMA = """\
121#if defined(__OPENCL_VERSION__) // && !defined(cl_khr_fp16)
122#  pragma OPENCL EXTENSION cl_khr_fp16: enable
123#endif
124"""
125
126_F64_PRAGMA = """\
127#if defined(__OPENCL_VERSION__) // && !defined(cl_khr_fp64)
128#  pragma OPENCL EXTENSION cl_khr_fp64: enable
129#endif
130"""
131
132def use_opencl():
133    return HAVE_OPENCL and os.environ.get("SAS_OPENCL", "").lower() != "none"
134
135ENV = None
136def reset_environment():
137    """
138    Call to create a new OpenCL context, such as after a change to SAS_OPENCL.
139    """
140    global ENV
141    ENV = GpuEnvironment() if use_opencl() else None
142
143def environment():
144    # type: () -> "GpuEnvironment"
145    """
146    Returns a singleton :class:`GpuEnvironment`.
147
148    This provides an OpenCL context and one queue per device.
149    """
150    if ENV is None:
151        if not HAVE_OPENCL:
152            raise RuntimeError("OpenCL startup failed with ***"
153                               + OPENCL_ERROR + "***; using C compiler instead")
154        reset_environment()
155        if ENV is None:
156            raise RuntimeError("SAS_OPENCL=None in environment")
157    return ENV
158
159def has_type(device, dtype):
160    # type: (cl.Device, np.dtype) -> bool
161    """
162    Return true if device supports the requested precision.
163    """
164    if dtype == generate.F32:
165        return True
166    elif dtype == generate.F64:
167        return "cl_khr_fp64" in device.extensions
168    elif dtype == generate.F16:
169        return "cl_khr_fp16" in device.extensions
170    else:
171        return False
172
173def get_warp(kernel, queue):
174    # type: (cl.Kernel, cl.CommandQueue) -> int
175    """
176    Return the size of an execution batch for *kernel* running on *queue*.
177    """
178    return kernel.get_work_group_info(
179        cl.kernel_work_group_info.PREFERRED_WORK_GROUP_SIZE_MULTIPLE,
180        queue.device)
181
182def _stretch_input(vector, dtype, extra=1e-3, boundary=32):
183    # type: (np.ndarray, np.dtype, float, int) -> np.ndarray
184    """
185    Stretch an input vector to the correct boundary.
186
187    Performance on the kernels can drop by a factor of two or more if the
188    number of values to compute does not fall on a nice power of two
189    boundary.   The trailing additional vector elements are given a
190    value of *extra*, and so f(*extra*) will be computed for each of
191    them.  The returned array will thus be a subset of the computed array.
192
193    *boundary* should be a power of 2 which is at least 32 for good
194    performance on current platforms (as of Jan 2015).  It should
195    probably be the max of get_warp(kernel,queue) and
196    device.min_data_type_align_size//4.
197    """
198    remainder = vector.size % boundary
199    if remainder != 0:
200        size = vector.size + (boundary - remainder)
201        vector = np.hstack((vector, [extra] * (size - vector.size)))
202    return np.ascontiguousarray(vector, dtype=dtype)
203
204
205def compile_model(context, source, dtype, fast=False):
206    # type: (cl.Context, str, np.dtype, bool) -> cl.Program
207    """
208    Build a model to run on the gpu.
209
210    Returns the compiled program and its type.
211
212    Raises an error if the desired precision is not available.
213    """
214    dtype = np.dtype(dtype)
215    if not all(has_type(d, dtype) for d in context.devices):
216        raise RuntimeError("%s not supported for devices"%dtype)
217
218    source_list = [generate.convert_type(source, dtype)]
219
220    if dtype == generate.F16:
221        source_list.insert(0, _F16_PRAGMA)
222    elif dtype == generate.F64:
223        source_list.insert(0, _F64_PRAGMA)
224
225    # Note: USE_SINCOS makes the intel cpu slower under opencl
226    if context.devices[0].type == cl.device_type.GPU:
227        source_list.insert(0, "#define USE_SINCOS\n")
228    options = (get_fast_inaccurate_build_options(context.devices[0])
229               if fast else [])
230    source = "\n".join(source_list)
231    program = cl.Program(context, source).build(options=options)
232    #print("done with "+program)
233    return program
234
235
236# for now, this returns one device in the context
237# TODO: create a context that contains all devices on all platforms
238class GpuEnvironment(object):
239    """
240    GPU context, with possibly many devices, and one queue per device.
241    """
242    def __init__(self):
243        # type: () -> None
244        # find gpu context
245        #self.context = cl.create_some_context()
246
247        self.context = None
248        if 'SAS_OPENCL' in os.environ:
249            #Setting PYOPENCL_CTX as a SAS_OPENCL to create cl context
250            os.environ["PYOPENCL_CTX"] = os.environ["SAS_OPENCL"]
251        if 'PYOPENCL_CTX' in os.environ:
252            self._create_some_context()
253
254        if not self.context:
255            self.context = _get_default_context()
256
257        # Byte boundary for data alignment
258        #self.data_boundary = max(d.min_data_type_align_size
259        #                         for d in self.context.devices)
260        self.queues = [cl.CommandQueue(context, context.devices[0])
261                       for context in self.context]
262        self.compiled = {}
263
264    def has_type(self, dtype):
265        # type: (np.dtype) -> bool
266        """
267        Return True if all devices support a given type.
268        """
269        return any(has_type(d, dtype)
270                   for context in self.context
271                   for d in context.devices)
272
273    def get_queue(self, dtype):
274        # type: (np.dtype) -> cl.CommandQueue
275        """
276        Return a command queue for the kernels of type dtype.
277        """
278        for context, queue in zip(self.context, self.queues):
279            if all(has_type(d, dtype) for d in context.devices):
280                return queue
281
282    def get_context(self, dtype):
283        # type: (np.dtype) -> cl.Context
284        """
285        Return a OpenCL context for the kernels of type dtype.
286        """
287        for context in self.context:
288            if all(has_type(d, dtype) for d in context.devices):
289                return context
290
291    def _create_some_context(self):
292        # type: () -> cl.Context
293        """
294        Protected call to cl.create_some_context without interactivity.  Use
295        this if SAS_OPENCL is set in the environment.  Sets the *context*
296        attribute.
297        """
298        try:
299            self.context = [cl.create_some_context(interactive=False)]
300        except Exception as exc:
301            warnings.warn(str(exc))
302            warnings.warn("pyopencl.create_some_context() failed")
303            warnings.warn("the environment variable 'SAS_OPENCL' might not be set correctly")
304
305    def compile_program(self, name, source, dtype, fast, timestamp):
306        # type: (str, str, np.dtype, bool, float) -> cl.Program
307        """
308        Compile the program for the device in the given context.
309        """
310        # Note: PyOpenCL caches based on md5 hash of source, options and device
311        # so we don't really need to cache things for ourselves.  I'll do so
312        # anyway just to save some data munging time.
313        tag = generate.tag_source(source)
314        key = "%s-%s-%s%s"%(name, dtype, tag, ("-fast" if fast else ""))
315        # Check timestamp on program
316        program, program_timestamp = self.compiled.get(key, (None, np.inf))
317        if program_timestamp < timestamp:
318            del self.compiled[key]
319        if key not in self.compiled:
320            context = self.get_context(dtype)
321            logging.info("building %s for OpenCL %s", key,
322                         context.devices[0].name.strip())
323            program = compile_model(self.get_context(dtype),
324                                    str(source), dtype, fast)
325            self.compiled[key] = (program, timestamp)
326        return program
327
328def _get_default_context():
329    # type: () -> List[cl.Context]
330    """
331    Get an OpenCL context, preferring GPU over CPU, and preferring Intel
332    drivers over AMD drivers.
333    """
334    # Note: on mobile devices there is automatic clock scaling if either the
335    # CPU or the GPU is underutilized; probably doesn't affect us, but we if
336    # it did, it would mean that putting a busy loop on the CPU while the GPU
337    # is running may increase throughput.
338    #
339    # Macbook pro, base install:
340    #     {'Apple': [Intel CPU, NVIDIA GPU]}
341    # Macbook pro, base install:
342    #     {'Apple': [Intel CPU, Intel GPU]}
343    # 2 x nvidia 295 with Intel and NVIDIA opencl drivers installed
344    #     {'Intel': [CPU], 'NVIDIA': [GPU, GPU, GPU, GPU]}
345    gpu, cpu = None, None
346    for platform in cl.get_platforms():
347        # AMD provides a much weaker CPU driver than Intel/Apple, so avoid it.
348        # If someone has bothered to install the AMD/NVIDIA drivers, prefer
349        # them over the integrated graphics driver that may have been supplied
350        # with the CPU chipset.
351        preferred_cpu = (platform.vendor.startswith('Intel')
352                         or platform.vendor.startswith('Apple'))
353        preferred_gpu = (platform.vendor.startswith('Advanced')
354                         or platform.vendor.startswith('NVIDIA'))
355        for device in platform.get_devices():
356            if device.type == cl.device_type.GPU:
357                # If the existing type is not GPU then it will be CUSTOM
358                # or ACCELERATOR so don't override it.
359                if gpu is None or (preferred_gpu and gpu.type == cl.device_type.GPU):
360                    gpu = device
361            elif device.type == cl.device_type.CPU:
362                if cpu is None or preferred_cpu:
363                    cpu = device
364            else:
365                # System has cl.device_type.ACCELERATOR or cl.device_type.CUSTOM
366                # Intel Phi for example registers as an accelerator
367                # Since the user installed a custom device on their system
368                # and went through the pain of sorting out OpenCL drivers for
369                # it, lets assume they really do want to use it as their
370                # primary compute device.
371                gpu = device
372
373    # order the devices by gpu then by cpu; when searching for an available
374    # device by data type they will be checked in this order, which means
375    # that if the gpu supports double then the cpu will never be used (though
376    # we may make it possible to explicitly request the cpu at some point).
377    devices = []
378    if gpu is not None:
379        devices.append(gpu)
380    if cpu is not None:
381        devices.append(cpu)
382    return [cl.Context([d]) for d in devices]
383
384
385class GpuModel(KernelModel):
386    """
387    GPU wrapper for a single model.
388
389    *source* and *model_info* are the model source and interface as returned
390    from :func:`generate.make_source` and :func:`generate.make_model_info`.
391
392    *dtype* is the desired model precision.  Any numpy dtype for single
393    or double precision floats will do, such as 'f', 'float32' or 'single'
394    for single and 'd', 'float64' or 'double' for double.  Double precision
395    is an optional extension which may not be available on all devices.
396    Half precision ('float16','half') may be available on some devices.
397    Fast precision ('fast') is a loose version of single precision, indicating
398    that the compiler is allowed to take shortcuts.
399    """
400    def __init__(self, source, model_info, dtype=generate.F32, fast=False):
401        # type: (Dict[str,str], ModelInfo, np.dtype, bool) -> None
402        self.info = model_info
403        self.source = source
404        self.dtype = dtype
405        self.fast = fast
406        self.program = None # delay program creation
407        self._kernels = None
408
409    def __getstate__(self):
410        # type: () -> Tuple[ModelInfo, str, np.dtype, bool]
411        return self.info, self.source, self.dtype, self.fast
412
413    def __setstate__(self, state):
414        # type: (Tuple[ModelInfo, str, np.dtype, bool]) -> None
415        self.info, self.source, self.dtype, self.fast = state
416        self.program = None
417
418    def make_kernel(self, q_vectors):
419        # type: (List[np.ndarray]) -> "GpuKernel"
420        if self.program is None:
421            compile_program = environment().compile_program
422            with open('model.c','w') as fid:
423                print(self.source['opencl'], file=fid)
424            timestamp = generate.ocl_timestamp(self.info)
425            self.program = compile_program(
426                self.info.name,
427                self.source['opencl'],
428                self.dtype,
429                self.fast,
430                timestamp)
431            variants = ['Iq', 'Iqxy', 'Imagnetic']
432            names = [generate.kernel_name(self.info, k) for k in variants]
433            kernels = [getattr(self.program, k) for k in names]
434            self._kernels = dict((k, v) for k, v in zip(variants, kernels))
435        is_2d = len(q_vectors) == 2
436        if is_2d:
437            kernel = [self._kernels['Iqxy'], self._kernels['Imagnetic']]
438        else:
439            kernel = [self._kernels['Iq']]*2
440        return GpuKernel(kernel, self.dtype, self.info, q_vectors)
441
442    def release(self):
443        # type: () -> None
444        """
445        Free the resources associated with the model.
446        """
447        if self.program is not None:
448            self.program = None
449
450    def __del__(self):
451        # type: () -> None
452        self.release()
453
454# TODO: check that we don't need a destructor for buffers which go out of scope
455class GpuInput(object):
456    """
457    Make q data available to the gpu.
458
459    *q_vectors* is a list of q vectors, which will be *[q]* for 1-D data,
460    and *[qx, qy]* for 2-D data.  Internally, the vectors will be reallocated
461    to get the best performance on OpenCL, which may involve shifting and
462    stretching the array to better match the memory architecture.  Additional
463    points will be evaluated with *q=1e-3*.
464
465    *dtype* is the data type for the q vectors. The data type should be
466    set to match that of the kernel, which is an attribute of
467    :class:`GpuProgram`.  Note that not all kernels support double
468    precision, so even if the program was created for double precision,
469    the *GpuProgram.dtype* may be single precision.
470
471    Call :meth:`release` when complete.  Even if not called directly, the
472    buffer will be released when the data object is freed.
473    """
474    def __init__(self, q_vectors, dtype=generate.F32):
475        # type: (List[np.ndarray], np.dtype) -> None
476        # TODO: do we ever need double precision q?
477        env = environment()
478        self.nq = q_vectors[0].size
479        self.dtype = np.dtype(dtype)
480        self.is_2d = (len(q_vectors) == 2)
481        # TODO: stretch input based on get_warp()
482        # not doing it now since warp depends on kernel, which is not known
483        # at this point, so instead using 32, which is good on the set of
484        # architectures tested so far.
485        if self.is_2d:
486            # Note: 16 rather than 15 because result is 1 longer than input.
487            width = ((self.nq+16)//16)*16
488            self.q = np.empty((width, 2), dtype=dtype)
489            self.q[:self.nq, 0] = q_vectors[0]
490            self.q[:self.nq, 1] = q_vectors[1]
491        else:
492            # Note: 32 rather than 31 because result is 1 longer than input.
493            width = ((self.nq+32)//32)*32
494            self.q = np.empty(width, dtype=dtype)
495            self.q[:self.nq] = q_vectors[0]
496        self.global_size = [self.q.shape[0]]
497        context = env.get_context(self.dtype)
498        #print("creating inputs of size", self.global_size)
499        self.q_b = cl.Buffer(context, mf.READ_ONLY | mf.COPY_HOST_PTR,
500                             hostbuf=self.q)
501
502    def release(self):
503        # type: () -> None
504        """
505        Free the memory.
506        """
507        if self.q_b is not None:
508            self.q_b.release()
509            self.q_b = None
510
511    def __del__(self):
512        # type: () -> None
513        self.release()
514
515class GpuKernel(Kernel):
516    """
517    Callable SAS kernel.
518
519    *kernel* is the GpuKernel object to call
520
521    *model_info* is the module information
522
523    *q_vectors* is the q vectors at which the kernel should be evaluated
524
525    *dtype* is the kernel precision
526
527    The resulting call method takes the *pars*, a list of values for
528    the fixed parameters to the kernel, and *pd_pars*, a list of (value,weight)
529    vectors for the polydisperse parameters.  *cutoff* determines the
530    integration limits: any points with combined weight less than *cutoff*
531    will not be calculated.
532
533    Call :meth:`release` when done with the kernel instance.
534    """
535    def __init__(self, kernel, dtype, model_info, q_vectors):
536        # type: (cl.Kernel, np.dtype, ModelInfo, List[np.ndarray]) -> None
537        q_input = GpuInput(q_vectors, dtype)
538        self.kernel = kernel
539        self.info = model_info
540        self.dtype = dtype
541        self.dim = '2d' if q_input.is_2d else '1d'
542        # plus three for the normalization values
543        self.result = np.empty(2*q_input.nq+2,dtype)
544
545        # Inputs and outputs for each kernel call
546        # Note: res may be shorter than res_b if global_size != nq
547        env = environment()
548        self.queue = env.get_queue(dtype)
549
550        self.result_b = cl.Buffer(self.queue.context, mf.READ_WRITE,
551                                  q_input.global_size[0] * dtype.itemsize)
552        self.q_input = q_input # allocated by GpuInput above
553
554        self._need_release = [self.result_b, self.q_input]
555        self.real = (np.float32 if dtype == generate.F32
556                     else np.float64 if dtype == generate.F64
557                     else np.float16 if dtype == generate.F16
558                     else np.float32)  # will never get here, so use np.float32
559    __call__= Iq
560
561    def Iq(self, call_details, values, cutoff, magnetic):
562        # type: (CallDetails, np.ndarray, np.ndarray, float, bool) -> np.ndarray
563        context = self.queue.context
564        # Arrange data transfer to card
565        details_b = cl.Buffer(context, mf.READ_ONLY | mf.COPY_HOST_PTR,
566                              hostbuf=call_details.buffer)
567        values_b = cl.Buffer(context, mf.READ_ONLY | mf.COPY_HOST_PTR,
568                             hostbuf=values)
569
570        kernel = self.kernel[1 if magnetic else 0]
571        args = [
572            np.uint32(self.q_input.nq), None, None,
573            details_b, values_b, self.q_input.q_b, self.result_b,
574            self.real(cutoff),
575        ]
576        #print("Calling OpenCL")
577        call_details.show(values)
578        #Call kernel and retrieve results
579        wait_for = None
580        last_nap = time.clock()
581        step = 1000000//self.q_input.nq + 1
582        for start in range(0, call_details.num_eval, step):
583            stop = min(start + step, call_details.num_eval)
584            #print("queuing",start,stop)
585            args[1:3] = [np.int32(start), np.int32(stop)]
586            wait_for = [kernel(self.queue, self.q_input.global_size, None,
587                               *args, wait_for=wait_for)]
588            if stop < call_details.num_eval:
589                # Allow other processes to run
590                wait_for[0].wait()
591                current_time = time.clock()
592                if current_time - last_nap > 0.5:
593                    time.sleep(0.05)
594                    last_nap = current_time
595        cl.enqueue_copy(self.queue, self.result, self.result_b)
596        #print("result", self.result)
597
598        # Free buffers
599        for v in (details_b, values_b):
600            if v is not None:
601                v.release()
602
603        pd_norm = self.result[self.q_input.nq]
604        scale = values[0]/(pd_norm if pd_norm != 0.0 else 1.0)
605        background = values[1]
606        #print("scale",scale,values[0],self.result[self.q_input.nq],background)
607        return scale*self.result[:self.q_input.nq] + background
608        # return self.result[:self.q_input.nq]
609     #NEEDS TO BE FINISHED FOR OPENCL
610     def beta():
611         return 0
612
613    def release(self):
614        # type: () -> None
615        """
616        Release resources associated with the kernel.
617        """
618        for v in self._need_release:
619            v.release()
620        self._need_release = []
621
622    def __del__(self):
623        # type: () -> None
624        self.release()
Note: See TracBrowser for help on using the repository browser.